GPU glossary
Plain-language definitions of the terms used on spec sheets, from CUDA cores and Infinity Cache to memory bandwidth and TOPS.
- CUDA cores
- NVIDIA's name for its FP32 shader ALUs. Counts are only comparable within one architecture.
- SM (Streaming Multiprocessor)
- The repeating compute block in an NVIDIA GPU. Since Ampere each SM has 128 CUDA cores, 4 Tensor Cores and 1 RT core.
- Tensor Cores
- Matrix-math units used for AI and DLSS. Blackwell's 5th-generation Tensor Cores add FP4.
- RT cores
- Fixed-function units that accelerate ray-triangle intersection for ray and path tracing.
- AI TOPS
- NVIDIA's headline AI figure for GeForce and RTX PRO. For RTX 50 it is FP4 with sparsity; for RTX 40 it is FP8 with sparsity. Do not compare across generations.
- TGP / TBP
- Total graphics (or board) power: the power limit of the card. For laptops the range is set by the laptop maker.
- Memory bandwidth
- Bus width × per-pin data rate ÷ 8. A 256-bit bus at 28 Gbps gives 896 GB/s.
- GDDR7 / GDDR6X / HBM
- Graphics memory types. HBM (High Bandwidth Memory) is stacked beside the GPU and used in data-center parts.
- NVFP4
- NVIDIA's 4-bit floating-point format for AI inference, introduced with Blackwell.
- NVLink
- NVIDIA's GPU-to-GPU interconnect. NVL72 means 72 GPU packages share one NVLink domain.
- Superchip
- A board or package combining NVIDIA's own CPU (Grace or Vera) with one or more GPUs over NVLink-C2C.
- MIG
- Multi-Instance GPU: splitting one GPU into isolated partitions.
- DLSS
- Deep Learning Super Sampling. DLSS 4 added Multi Frame Generation; DLSS 4.5 added Dynamic MFG; DLSS 5 (2026) adds 3D-Guided Neural Rendering on RTX 50.
- Stream processors / CUs (AMD)
- AMD's shader ALUs, grouped into compute units of 64. RDNA 3 and 4 can dual-issue, so AMD's FP32 figures are roughly double a simple core × clock count.
- Infinity Cache
- AMD's large on-die last-level cache, which raises effective memory bandwidth.
- Regional launch price
- The price NVIDIA or AMD published for a country at launch, including local sales tax where that market quotes it. Prices differ from a straight currency conversion because of tax, import duty and warranty rules.
- Xe-cores and XMX (Intel)
- Intel groups its shader ALUs into Xe-cores of 128; XMX engines are its matrix units for AI and XeSS. Intel quotes AI throughput as INT8 TOPS.